No-Speech Probability
A model-produced probability or confidence signal indicating that an audio segment does not contain meaningful speech.
Separating silence from speech with an energy threshold alone is unreliable across recording conditions. Some ASR models expose a dedicated no-speech score during decoding; combined with VAD, token confidence, and other quality signals, it can help suppress unnecessary transcription or hallucinated text.
One threshold should not be assumed to transfer across microphones, noise levels, and languages. Moving the cutoff directly trades missed speech against false speech. VAD, segment duration, and endpointing trade-offs in real-time ASR covers the same decision at stream level.